Bridge Input (FLOAT)

The INT bridge handles steps and seeds; Bridge Input (FLOAT) is for the knobs that need decimals - CFG scale, denoise strength, LoRA weight, clip skip values that aren't whole. When an AI agent is tuning your workflow run over run, these are often exactly what it wants to sweep: try CFG 6.5, then 7.2, see which sticks. This node is the socket for that.

Same contract as every ComfyBridgeIO node. Drop it in, give it a name, save the workflow as API format JSON, and the Python client patches its value at runtime - bridge.run("workflow.json", inputs={"cfg": 7.5}). The client deep-copies the workflow (your original file never changes), posts it to /prompt, waits on the websocket, collects results. The node is a passthrough that marks where the float enters the graph.

The inputs and output that matter

• name (STRING, default "input_name") - the identifier. It must exactly match the key in your client's inputs dict. cfg in the widget, {"cfg": 7.5} in Python. Mismatch → ValueError: BridgeInput 'x' not found.
• value (FLOAT, default 0.0, min −1e9, max 1e9) - the workflow's default when nothing is injected. Set it to a reasonable CFG (7.0-ish) or denoise (0.5) so the graph still runs if you hit Queue by hand.
• value_link (optional FLOAT socket) - an upstream float output that overrides both the widget and any injected value. Skip it unless you want a fixed source; it shadows the agent.

Output is a single value (FLOAT) socket - feed it to a KSampler's cfg or denoise, a LoRA loader's strength, or any node that wants a real number.

Install

ComfyUI Manager → search ComfyBridgeIO → Install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/cbpark84/ComfyBridgeIO

Restart. No model downloads, no heavy dependencies - the nodes run on stock ComfyUI. The Python client (optional, pip install -e in the repo) pulls in aiohttp, websockets, and Pillow. One stale line to dodge: the README's clone URL says nicekate/ComfyBridgeIO; the actual repo is cbpark84 as above.

Common issues

The usual suspects are the same as the INT variant: name doesn't match the dict key, or the value you send is an int and the client's float() cast works fine but you lose the decimal precision you were after (sending 7 becomes 7.0, harmless). The one real trap is value_link - wire it in once, forget it, and every agent sweep silently goes nowhere because the connected value wins.